fix transcribe bug

app.py

@@ -1,4 +1,4 @@
-# app.py – FINAL & WORKING on paid T4 (15–25 s for 3 min)
+# app.py – FIXED: no_timestamps_token_id added (no more ValueError)
 import os
 import gradio as gr
 import spaces
@@ -16,28 +16,31 @@ pipe = pipeline(
     token=os.getenv("HF_TOKEN")
 )
 
-# ←←← THIS FIXES THE lang_to_id ERROR FOREVER
+# Fix old Whisper config completely (including timestamps token)
 if not hasattr(pipe.model.generation_config, "lang_to_id"):
     pipe.model.generation_config.lang_to_id = {"is": 50259}
     pipe.model.generation_config.task_to_id = {"transcribe": 50359}
     pipe.model.generation_config.forced_decoder_ids = None
 
+# ←←← THIS FIXES THE TIMESTAMP ERROR
+if not hasattr(pipe.model.generation_config, "no_timestamps_token_id"):
+    pipe.model.generation_config.no_timestamps_token_id = 50363
+
 pipe.model.generation_config.language = "is"
 pipe.model.generation_config.task = "transcribe"
 
-print("Model ready – locked to Icelandic – no more errors!")
+print("Model ready – fully fixed for timestamps!")
 
 @spaces.GPU(duration=120)
 def transcribe_safe(audio_path):
     if not audio_path:
         return "Hladdu upp hljóðskrá"
     
-    # ← librosa imported here so startup never crashes
     import librosa
     
     audio, sr = librosa.load(audio_path, sr=16000)
-    chunk_len = 16000 * 100  # 20 seconds
-    stride    = 16000 * 2     # 2 seconds overlap
+    chunk_len = 16000 * 20
+    stride = 16000 * 2
     chunks = []
     for i in range(0, len(audio), chunk_len - stride):
         chunk = audio[i:i + chunk_len]
@@ -52,7 +55,6 @@ def transcribe_safe(audio_path):
     
     return full_text.strip() or "Ekkert heyrt"
 
-# Your original beautiful UI
 with gr.Blocks(title="Íslenskt ASR – 3 mín T4 Paid") as demo:
     gr.Markdown("# Íslenskt ASR – 3 mín hljóð")
     gr.Markdown("**~4 % WER · 15–25 sek · T4 Paid**")